Henry Markram :用超级计算机构造大脑

==========================
http://dotsub.com/view/85c7b35b-fa45-4224-a677-4079fdcb987b
Henry Markram :用超级计算机构造大脑
我们的任务是建立 一个复杂的，可行的 人脑计算机模型。 过去四年，我们已经 在一小块鼠脑上 进行概念验证， 根据这样的概念验证，我们正扩大该项目规模到 人脑大小。
我们为什么要这样做呢？ 有三个重要原因。 首先，了解人类的大脑对我们来说非常重要 如果我们要继续在社会中生存 同时我认为也是进化上的关键一步。 第二个原因是， 我们不能总是进行动物试验， 我们要把我们所有的数据和知识包含进 一个可用模型中。 好比诺亚方舟，好比一个档案馆。 第三个原因是：地球上有20亿人 患有精神疾病， 他们的用药 主要依靠经验。 我认为，我们能就如何对待疾病的 有非常具体的解决办法。
现在，即使在这个阶段， 我们可以利用大脑模型 探讨一些基本的问题 关于大脑是如何工作的。 这里，首次通过TED 我想和大家分享我们如何解决 一个理论 -- 有许多理论 -- 一个关于大脑如何工作的理论。 这种理论认为大脑 创建，构建了一个版本的宇宙。 把这个宇宙作映射， 向气泡一样，映射到我们周围。
当然，这是数百年哲学辩论的题目。 但是，有史以来，我们实际上可以解决这个问题， 依靠大脑仿真， 同时提出非常系统和严谨的问题， 这个理论是否有可能是正确的。 之所以月亮在地平线上是巨大的 纯粹是因为我们的知觉气泡 没有延伸到三十八万公里外。 用光了空间。 所以我们比较建筑物 在我们的知觉气泡范围内， 我们做出一个判断。 我们做出判断，那个建筑很大， 即使它并没有那么大，
这表明 判断是支持 知觉气泡的关键。判断维系了知觉气泡。 没有判断你不能看，不能思考， 不能感知。 没准你会认为麻醉药的工作原理是 让你熟睡， 或者阻断感知器官，使你感不到疼痛， 不过，实际上多数麻醉药不是这么工作的。 麻醉药引入干扰 到大脑中，这样神经元互相之间不能理解。 神经元糊涂了， 你也就不能做出一个判断。 所以，当你正试图弥补你的思绪的， 以为手术医生 正在做手术，实际上医生早走了。 正在家里喝茶呢。 （笑）
当你拾阶而上打开一扇门的时候， 你不由自主的 做出判断， 成千上万的判断，诸如房间的大小， 墙壁，高度，房内物体。 99%的你所见， 并非来自你的眼睛。 是你对这个房间推断。 所以我可以确定的说， “我思故我在。“ 但我不能说，“你思故你在。” 因为你包含在我的知觉气泡中。
我们可以就此进行哲学思考， 但我们没必要再思考上百年。 我们可以提出非常具体的问题。 “大脑能建立这样一个知觉吗？” 有能力做吗? 大脑有物质创造这些吗? 这就是我今天要讲的东西。
宇宙用了110亿年时间创造的大脑。 宇宙一点点的完善大脑。 增加了额叶，所以有了本能， 因为要适应陆地生活。 不过真正的进步是大脑皮层。 一个新大脑。你需要它。 哺乳动物需要它 因为哺乳动物需要适应父母的角色， 社会互动， 复杂的认知功能。
你可以认为大脑皮层 是现在最终解决方案， 就我们所知的宇宙。 大脑皮层是顶峰，是宇宙的 最终产品。 在进化上很成功 从老鼠到人类 扩展千倍的神经元细胞， 来制造这吓人的 器官，组织。 它的进化没有停止。 实际上，人脑中的大脑皮层 还在飞快的进化中。
如果你放大到大脑皮层表面， 你会发现它有许多小模块组成， 就像计算机里的G5处理器。 只不过，大脑皮层有上百万块。 这些模块在进化上非常成功 我们所作的 就是在大脑中不断的复制这些模块 直到我们用光了头骨中的空间。 大脑开始折叠自身， 这就是为什么大脑皮层如此褶皱。 我们只有成列的包装， 这样才能有更多的皮层 实现更复杂的功能。
你可以把大脑皮层想想为 一台巨大的钢琴， 由上百万个琴键的钢琴。 每一个皮层 发出一个音符。 你刺激它；大脑就奏出交响乐。 不只是知觉的交响乐。 是你的宇宙，你的现实的交响乐。 现在，当然需要很多年 才能掌握有上百万个琴键的钢琴。 这就是为什么你送你的孩子上好学校， 希望最终上牛津。 但不只是教育。 这也是遗传。 可能你生来幸运， 或者你知道如何掌握皮层。 你可以奏出美妙的交响乐。
实际上，有一个新的自闭症理论 叫做“紧张世界”理论， 提出皮层列是超级列。 它们超有活性，可塑性， 这样自闭症者可能有能力 建立并学习一首交响乐 而对普通人来说不可想象。 你也可以理解 如果你有一种疾病 在这些皮层列中， 音符关闭了。 知觉，也就是你创造的交响乐 也随之跑调， 你会有疾病的症状。
所以，神经科学的圣杯 就是理解皮层的设计 -- 不只对神经科学； 可能对理解知觉，理解现实有帮助， 还可能对理解物理现实又帮助。 所以过去15年里， 我们所作的就是系统的解刨大脑皮层。 就像是去分类一片雨林。 有多少树？ 树的形状？ 每种树有多少棵？在哪里？
不过可能比分类还要复杂，因为需要你 描述并发现所有的通讯规律， 连接的规律， 因为神经元不是随便和别的神经元连接的。 神经元挑选连接对象非常谨慎。 比分类还复杂的是 因为需要你建立三维 数字模型。 我们为数万个神经元 建立了数字模型，我所见到的 不同种类的神经元。 一旦你有了这些，你就可以 建立皮层列。
这里，我们把它们缠绕在一起。 这样做的时候，你所见的 就是分叉的相交 实践上再数百万的点上。 每个交叉 都形成一个突触 每个突触就是一个化学位置 这里可以进行相互通讯。 这些突触一起 构成了网络 或者说是大脑的电路 电路，也可以想象为 大脑的编制结构。 当你想到大脑的编制结构时， 结构，如何构建？这个编织物的图案是什么？ 你意识到这个构成 是对任何大脑理论的挑战， 特别对一种理论，说 现实浮现 出这个编织物， 以一种特有的图案。
原因是大脑重要的设计秘密是 多样性。 每个神经元都不同。 如同在森林中。每棵松树都不同。 可以有许多同类的书， 但是每棵松树都不一样。大脑也一样。 我大脑中的每个神经元都不相同， 我大脑中的神经元也不会和你的一样。 你的大脑中的神经元 以一样与众不同。 你可能有或多或少的神经元。 所以不可能 有同样的编制结构，同样的电路。
这里，我们如何建立现实 而又彼此可以理解的呢？ 我们没有必要推测。 现在我们可以看看这一千万突触。 我们可以看看编制结构。我们可以变更神经元。 我可以用不同的神经元。 我们可以把怕们摆放到不同的位置上， 朝向不同的方位。 我们可以用或多或少的神经元。 我们这么做了之后， 我们发现，电路却是变化了。 但是电路设计的模式没有变化。 所以，大脑的编制结构， 不管你的大脑或大或小， 神经元的种类不同， 神经元的形态不同， 我们实际上共享了 同样的编制结构。 我们认为这是物种特异性的， 这意味着可以解释 为什么我们不能和别的物种交流。
好，打开它。你必须作的是 使它活起来。 我们让它们活过来 用数学公式，很多的数学运算。 实际上，这些让神经元变成发电机的数学公式 是被两个剑桥的诺贝尔奖得主发现的。 我们又让神经元活过来的数学。 我们也有 神经元采集信息的数学描述， 以及它们如何制造闪电 来和彼此通讯的。 当到达突触后， 它们有效的， 正确的，震动突触。 像是电击 从这些突触中释放化学物质。
我们有数学工具描述这一过程。 我们可以描述神经元之间的彼此通讯。 真正的少数 数学公式用来仿真 大脑皮层的活动。 不过你得要一个非常大的计算机。 实际上一台笔记本电脑 只够计算一个神经元用的。 所以得要一万台笔记本电脑。 所以需要IBM帮助， 那里有超级计算机， 可以把一万台笔记本电脑压缩到一个冰箱大笑。 所以我们有蓝色基因超级计算机。 我们可以加载所有的神经元， 每个在一个处理器上， 然后运行，看看发生了什么。 骑上魔毯。
我们激活它。看一看 当有刺激的时候 我们的大脑在做什么。 这是第一次看到。 当你第一次看到的时候，你会想， “老天。现实就从这里面产生的？“ 实际上，你可以开始 即使我们没有训练这个皮层列 来创建一个特定现实。 不过我们可以问，“玫瑰在哪里？” 我们可以问，"在那里， 如果我们用一幅图来刺激它？“ 在大脑皮层哪里？ 如果我们刺激它，最终它会在那里。
我们看待的方法 就是忽略神经元，忽略突触， 只看原始的电活动。 因为这就是它创建的。 创建电模式。 当我们这么做时， 我们实际上第一次看到， 看到这些幽灵一样的结构： 电物体出现 在皮层列中。 这些电物体 控制了所有刺激它的 信息。 当我们放大的时候， 确实像个宇宙。
下面一步是 提出这些大脑中的坐标 投射的知觉空间中。 如果你这样做， 你就可以走进 所创造的现实中 由这个机器创造的， 通过这一块大脑。 总的来说， 我认为宇宙可以 可能 演化出能看到自己的大脑， 这是意识到自身的第一步。 还有很多需要做的来测试这个理论， 以及其他理论。 不过我希望你可以至少部分的相信 建造大脑不是不可能的。 我们可以在10年内完成， 如果我们成功了， 10年内我们会发到TED， 一个全息图。
谢谢。 （掌声）

-----------------
Henry Markram builds a brain in a supercomputer
Our mission is to build a detailed, realistic, computer model of the human brain. And we've done, in the past four years, a proof of concept, on a small part of the rodent brain, and with this proof of concept we are now scaling the project up to reach the human brain.

Why are we doing this? There are three important reasons. The first is, it's essential for us to understand the human brain if we do want to get along in society, and I think that it is a key step in evolution. The second reason is, we can not keep doing animal experimentation forever, and we have to embody all our data and all our knowledge, into a working model. It's like a Noah's Ark. It's like an archive. And the third reason is that there are two billion people on the planet that are affected by mental disorder, and the drugs that are used today are largely empirical. I think that we can come up with very concrete solutions on how to treat disorders.

Now, even at this stage, we can use the brain model to explore some fundamental questions about how the brain works. And here, at TED, for the first time, I'd like to share with you how we're addressing one theory -- there are many theories -- one theory of how the brain works. So, this theory is that the brain creates, builds, a version of the universe. And projects this version of the universe, like a bubble, all around us.

Now, this is of course a topic of philosophical debate for centuries. But, for the first time, we can actually address this, with brain simulation, and ask very systematic and rigorous questions, whether this theory could possibly be true. The reason why the moon is huge on the horizon is simply because our perceptual bubble does not stretch out 380,000 kilometers. It runs out of space. And so what we do is we compare the buildings within our perceptual bubble, and we make a decision. We make a decision it's that big, even though it's not that big,

and what that illustrates is that decisions are the key things that support our perceptual bubble. It keeps it alive. Without decisions you cannot see, you cannot think, you cannot feel. And you may think that anesthetics work by sending you into some deep sleep, or by blocking your receptors so that you don't feel pain, but in fact most anesthetics don't work that way. What they do is they introduce a noise into the brain so that the neurons can not understand each other. They are confused, and you can not make a decision. So, while you're trying to make up your mind, what the doctor, the surgeon is doing while he's hacking away at your body, he's long gone. He's at home having tea. (Laughter)

So, when you walk up to a door and you open it, what you compulsively have to do to perceive is to make decisions, thousands of decisions about the size of the room, the wall, the height, the objects in this room. 99 percent of what you see is not what comes in through the eyes. It is what you infer about that room. So I can say, with some certainty, "I think, therefore I am." But I cannot say, "You think, therefore you are," because "you" are within my perceptual bubble.

Now, we can speculate and philosophize this, but we don't actually have to for the next hundred years. We can ask a very concrete question. "Can the brain build such a perception?" Is it capable of doing it? Does it have the substance to do it? And that's what I'm going to describe to you today.

So, it took the universe 11 billion years to build the brain. It had to improve it a little bit. It had to add to the frontal part, so that you would have instincts, because they had to cope on land. But the real big step was the neocortex. It's a new brain. You needed it. The mammals needed it because they had to cope with parenthood, social interactions, complex cognitive functions.

So, you can think of the neocortex actually as the ultimate solution today, of the universe as we know it. It's the pinnacle, it's the final product that the universe has produced. It was so successful in evolution that from mouse to man it expanded about a thousandfold in terms of the numbers of neurons, to produce this almost frightening organ, structure. And it has not stopped its evolutionary path. In fact, the neocortex in the human brain is evolving at an enormous speed.

If you zoom into the surface of the neocortex, you discover that it's made up of little modules, G5 processors, like in a computer. But there are about a million of them. They were so successful in evolution that what we did was to duplicate them over and over and add more and more of them to the brain until we ran out of space in the skull. And the brain started to fold in on itself, and that's why the neocortex is so highly convoluted. We're just packing in columns, so that we'd have more neocortical columns to perform more complex functions.

So you can think of the neocortex actually as a massive grand piano, a million-key grand piano. Each of these neocortical columns would produce a note. You stimulate it; it produces a symphony. But it's not just a symphony of perception. It's a symphony of your universe, your reality. Now, of course it takes years to learn how to master a grand piano with a million keys. That's why you have to send your kids to good schools, hopefully eventually to Oxford. But it's not only education. It's also genetics. You may be born lucky, or you know how to master your neocortical column, and you can play a fantastic symphony.

In fact, there is a new theory of autism called the "intense world" theory, which suggests that the neocortical columns are super-columns. They are highly reactive, and they are super-plastic, and so the autists are probably capable of building and learning a symphony which is unthinkable for us. But you can also understand that if you have a disease within one of these columns, the note is going to be off. The perception, the symphony that you create is going to be corrupted, and you will have symptoms of disease.

So, the Holy Grail for neuroscience is really to understand the design of the neocoritical column -- and it's not just for neuroscience; it's perhaps to understand perception, to understand reality, and perhaps to even also understand physical reality. So, what we did was, for the past 15 years, was to dissect out the neocortex, systematically. It's a bit like going and cataloging a piece of the rainforest. How many trees does it have? What shapes are the trees? How many of each type of tree do you have? Where are they positioned?

But it's a bit more than cataloging because you actually have to describe and discover all the rules of communication, the rules of connectivity, because the neurons don't just like to connect with any neuron. They choose very carefully who they connect with. It's also more than cataloging because you actually have to build three-dimensional digital models of them. And we did that for tens of thousands of neurons, built digital models of all the different types of neurons we came across. And once you have that, you can actually begin to build the neocortical column.

And here we're coiling them up. But as you do this, what you see is that the branches intersect actually in millions of locations. and at each of these intersections they can form a synapse. And a synapse is a chemical location where they communicate with each other. And these synapses together form the network or the circuit of the brain. Now, the circuit, you could also think of as the fabric of the brain. And when you think of the fabric of the brain, the structure, how is it built? What is the pattern of the carpet? You realize that this poses a fundamental challenge to any theory of the brain, and especially to a theory that says that there is some reality that emerges out of this carpet, out of this particular carpet with a particular pattern.

The reason is because the most important design secret of the brain is diversity. Every neuron is different. It's the same in the forest. Every pine tree is different. You may have many different types of trees, but every pine tree is different. And in the brain it's the same. So there is no neuron in my brain that is the same as another, and there is no neuron in my brain that is the same as in yours. And your neurons are not going to be oriented and positioned in exactly the same way. And you may have more or less neurons. So it's very unlikely that you got the same fabric, the same circuitry.

So, how could we possibly create a reality that we can even understand each other? Well, we don't have to speculate. We can look at all 10 million synapses now. We can look at the fabric. And we can change neurons. We can use different neurons with different variations. We can position them in different places, orient them in different places. We can use less or more of them. And when we do that what we discovered is that the circuitry does change. But the pattern of how the circuitry is designed does not. So, the fabric of the brain, even though your brain may be smaller, bigger, it may have different types of neurons, different morphologies of neurons, we actually do share the same fabric. And we think this is species-specific, which means that that could explain why we can't communicate across species.

So, let's switch it on. But to do it, what you have to do is you have to make this come alive. We make it come alive with equations, a lot of mathematics. And, in fact, the equations that make neurons into electrical generators were discovered by two Cambridge Nobel Laureates. So, we have the mathematics to make the neurons come alive. We also have the mathematics to describe how neurons collect information, and how they create a little lightning bolt to communicate with each other. And when they get to the synapse, what they do is they effectively, literally, shock the synapse. It's like electrical shock that releases the chemicals from these synapses.

And we've got the mathematics to describe this process. So we can describe the communication between the neurons. There literally are only a handful of equations that you need to simulate the activity of the neocortex. But what you do need is a very big computer. And in fact you need one laptop to do all the calculations just for one neuron. So you need 10,000 laptops. So where do you go? You go to IBM, and you get a supercomputer, because they know how to take 10,000 laptops and put it into the size of a refrigerator. So now we have this Blue Gene supercomputer. We can load up all the neurons, each one on to its processor, and fire it up, and see what happens. Take the magic carpet for a ride.

Here we activate it. And this gives the first glimpse of what is happening in your brain when there is a stimulation. It's the first view. Now, when you look at that the first time, you think, "My god. How is reality coming out of that?" But, in fact, you can start, even though we haven't trained this neocortical column to create a specific reality. But we can ask, "Where is the rose?" We can ask, "Where is it inside, if we stimulate it with a picture?" Where is it inside the neocortex? Ultimately it's got to be there if we stimulated it with it.

So, the way that we can look at that is to ignore the neurons, ignore the synapses, and look just at the raw electrical activity. Because that is what it's creating. It's creating electrical patterns. So when we did this, we indeed, for the first time, saw these ghost-like structures: electrical objects appearing within the neocortical column. And it's these electrical objects that are holding all the information about whatever stimulated it. And then when we zoomed into this, it's like a veritable universe.

So the next step is just to take these brain coordinates and to project them into perceptual space. And if you do that, you will be able to step inside the reality that is created by this machine, by this piece of the brain. So, in summary, I think that the universe may have -- it's possible -- evolved a brain to see itself, which may be a first step in becoming aware of itself. There is a lot more to do to test these theories, and to test any other theories. But I hope that you are at least partly convinced that it is not impossible to build a brain. We can do it within 10 years, and if we do succeed, we will send to TED, in 10 years, a hologram to talk to you. Thank you. (Applause)